Fixed-Point Code Synthesis For Neural Networks
This work addresses the need for efficient neural network implementation in resource-constrained safety-critical systems, such as robots and autonomous vehicles, by providing a method to convert floating-point networks to fixed-point arithmetic, which is incremental as it builds on existing training and fixed-point optimization approaches.
The paper tackles the problem of deploying neural networks in safety-critical systems with limited computing resources by introducing a technique to tune fixed-point formats for already trained networks, ensuring output accuracy up to a user-defined threshold without modifying the original accuracy, as validated by experimental results showing the fixed-point network behaves identically to the floating-point version.
Over the last few years, neural networks have started penetrating safety critical systems to take decisions in robots, rockets, autonomous driving car, etc. A problem is that these critical systems often have limited computing resources. Often, they use the fixed-point arithmetic for its many advantages (rapidity, compatibility with small memory devices.) In this article, a new technique is introduced to tune the formats (precision) of already trained neural networks using fixed-point arithmetic, which can be implemented using integer operations only. The new optimized neural network computes the output with fixed-point numbers without modifying the accuracy up to a threshold fixed by the user. A fixed-point code is synthesized for the new optimized neural network ensuring the respect of the threshold for any input vector belonging the range [xmin, xmax] determined during the analysis. From a technical point of view, we do a preliminary analysis of our floating neural network to determine the worst cases, then we generate a system of linear constraints among integer variables that we can solve by linear programming. The solution of this system is the new fixed-point format of each neuron. The experimental results obtained show the efficiency of our method which can ensure that the new fixed-point neural network has the same behavior as the initial floating-point neural network.